Smart residential buildings are a critical component of contemporary urban development because building-level thermal regulation directly affects energy efficiency, occupant comfort, and the operational intelligence of wider smart-city infrastructure. This paper presents an attention-based long short-term memory fast model predictive control (ALSTM–FMPC) framework for thermal regulation in smart residential buildings. The method integrates sequence learning and predictive optimisation to address the limitations of conventional heating, ventilation, and cooling (HVAC) control systems that rely on fixed setpoints and static rules. The proposed controller was implemented on real-world data collected from a building automation system and was evaluated through profile-based comparative analysis. In the reported implementation, the baseline bench temperature profile ranges from 20 to 60∘C, the corresponding setpoints range from 20 to 55∘C, and HVAC outputs range from 0 to 80∘C. After applying ALSTM–FMPC, the temperature profile shifts to 35–75∘C, the setpoint span becomes 35–60∘C, and the controller signal extends to 0–100∘C. Comparative tests on the same bench show that the proposed controller outperforms classical LSTM and conventional model predictive control in responsiveness and adaptability. The base ALSTM configuration uses 150 recurrent hidden neurons, 100 non-recurrent hidden neurons, and 192,301 trainable parameters, while computational demand remains moderate, with memory usage in the 1.6–2.2 GB range and disk usage of 24.4 GB. These findings support the relevance of ALSTM–FMPC as an intelligent building-control strategy for smart-city residential energy systems.